Data Visualization: Day 3

Erik Westlund

2025-06-12

Day 3: What to Visualize – From Question to Plot

Overview for Day 3

Scientific questions and DAGs
What to visualize: distributions, time, space, relationships, models
Key design principles: honesty, annotation, small multiples
Case studies: 7 applied examples

Theme: Science Drives Visualization

Data visualization is valuable for its scientific application
It helps us:
- Understand our data
- Understand our models
- Communicate our findings in an impactful way

DAGs, Scientific Complexity, Causal Inference, and Visualization

From Question to DAG to Plot

We’ll start with a scientific question.
We’ll sketch out the causal structure using a DAG.
From this, we can assess typical problems scientific studies like this face
We can then use the DAG to guide our visualization choices

Scientific Question

“What causes mothers to receive comprehensive postnatal care?”

Directed Acyclic Graphs

A DAG is a graph that represents the causal relationships between variables.
It is a directed graph, meaning that the edges have a direction.
It is an acyclic graph, meaning that there are no cycles in the graph.
It is a graph, meaning that it has nodes and edges.

How DAGs Help Us In Science & Visualization

DAGs help us:
- Identify the variables that are important to the question
- Identify the variables that are confounding
- Identify the variables that are missing
- Identify the variables that are measured incorrectly
Having a good idea of the causal structure of a question helps us know what to visualize.

DAGs as a Visualization Tool

DAGs:
1. are themselves visualizations
2. encode scientific problems and causal structures
We can use the daggity R package to visualize the the causal structure of the problems around the motivating research question

Big Picture

rcp_dag <- dagify(RCP ~ PQ + PC + WTP + PT + RA + RP)

rcp_dag |>
    ggdag() +
    theme_dag()

RCP = “Received Comprehensive Postnatal Care”
PQ = “Provider Quality”
PC = “Personal Capacity”
WTP = “Willingness to Pay”
PT = “Provider Trust”
RA = “Risk Aversion”
RP = “Risk Profile”

Provider Quality (PQ) Sub-DAG

pq_dag <- dagify(
  PQ ~ PEC,
  PEC ~ ST
)

pq_dag |>
    ggdag() +
    theme_dag()

PQ = “Provider Quality”
PEC = “Political/Economic Conditions”
ST = “State”

Personal Capacity (PC) Sub-DAG

pc_dag <- dagify(
  PC ~ DEP + JOB + INC + DIS,
  DEP ~ INC + JOB + CO,
  JOB ~ EDU + CO + R + M + I + CC,
  DIS ~ PEC,
  INC ~ JOB + EDU + RE + PEC + AGE
)

pc_dag |>
    ggdag() +
    theme_dag()

PC = “Personal Capacity”
DEP = “Dependents”
JOB = “Job Type”
INC = “Income”
DIS = “Distance to Provider”
CO = “Cultural Orientation”
EDU = “Educational Attainment”
R = “Resilience”
M = “Motivation”
I = “Intelligence”
CC = “Community Connections”
RE = “Race/Ethnicity”
AGE = “Age”

Willingness to Pay (WTP) Sub-DAG

wtp_dag <- dagify(
  WTP ~ PQ + INC + INS + RA + CO,
  INS ~ JOB + PEC + AGE,
  CO ~ PED + PIN + REL + PEC + CC
)

wtp_dag |>
    ggdag() +
    theme_dag()

WTP = “Willingness to Pay”
PQ = “Provider Quality”
INC = “Income”
INS = “Insurance”
RA = “Risk Aversion”
CO = “Cultural Orientation”
JOB = “Job Type”
PEC = “Political/Economic Conditions”
AGE = “Age”
PED = “Parent Education”
PIN = “Parent Income”
REL = “Religion”
CC = “Community Connections”

Provider Trust (PT) Sub-DAG

pt_dag <- dagify(
  PT ~ PQ + RE + CO,
  CO ~ PED + PIN + REL + PEC + CC
)

pt_dag |>
    ggdag() +
    theme_dag()

PT = “Provider Trust”
PQ = “Provider Quality”
RE = “Race/Ethnicity”
CO = “Cultural Orientation”
PED = “Parent Education”
PIN = “Parent Income”
REL = “Religion”
PEC = “Political/Economic Conditions”
CC = “Community Connections”

Risk Aversion (RA) Sub-DAG

ra_dag <- dagify(
  RA ~ PQ + RP + INS,
  RP ~ PQ + AGE + OBE + MG + DM + HD + PP + PR + HT + GHT,
  INS ~ JOB + PEC + AGE
)

ra_dag |>
    ggdag() +
    theme_dag()

RA = “Risk Aversion”
PQ = “Provider Quality”
RP = “Risk Profile”
INS = “Insurance”
AGE = “Age”
OBE = “Obesity”
MG = “Multiple Gestation”
DM = “Diabetes Mellitus”
HD = “Heart Disease”
PP = “Placenta Previa”
PR = “Preeclampsia”
HT = “Hypertension”
GHT = “Gestational Hypertension”
JOB = “Job Type”
PEC = “Political/Economic Conditions”

Risk Profile (RP) Sub-DAG

rp_dag <- dagify(
  RP ~ PQ + AGE + OBE + MG + DM + HD + PP + PR + HT + GHT,
  OBE ~ PEC + AGE,
  MG ~ AGE + OBE,
  DM ~ AGE + OBE + INC,
  HD ~ AGE + OBE + DM,
  PP ~ AGE + MG,
  PR ~ AGE + HT + GHT + MG,
  HT ~ AGE + OBE,
  GHT ~ HT + MG
)

rp_dag |>
    ggdag() +
    theme_dag()

RP = “Risk Profile”
PQ = “Provider Quality”
AGE = “Age”
OBE = “Obesity”
MG = “Multiple Gestation”
DM = “Diabetes Mellitus”
HD = “Heart Disease”
PP = “Placenta Previa”
PR = “Preeclampsia”
HT = “Hypertension”
GHT = “Gestational Hypertension”
PEC = “Political/Economic Conditions”
INC = “Income”

Complete DAG Codification

complete_dag <- dagify(
  RCP ~ PQ + PC + WTP + PT + RA + RP,
  PQ ~ PEC,
  PC ~ DEP + JOB + INC + DIS,
  WTP ~ PQ + INC + INS + RA + CO,
  PT ~ PQ + RE + CO,
  RA ~ PQ + RP + INS,
  RP ~ PQ + AGE + OBE + MG + DM + HD + PP + PR + HT + GHT,
  PEC ~ ST,
  DEP ~ INC + JOB + CO,
  JOB ~ EDU + CO + R + M + I + CC,
  DIS ~ PEC,
  INC ~ JOB + EDU + RE + PEC + AGE,
  INS ~ JOB + PEC + AGE,
  CO ~ PED + PIN + REL + PEC + CC,
  PIN ~ PEC + PED,
  EDU ~ PED + PIN + R + M + I + CC,
  R ~ PR,
  M ~ PM,
  I ~ PI,
  CC ~ PCC,
  REL ~ RE,
  OBE ~ PEC + AGE,
  MG ~ AGE + OBE,
  DM ~ AGE + OBE + INC,
  HD ~ AGE + OBE + DM,
  PP ~ AGE + MG,
  PR ~ AGE + HT + GHT + MG,
  HT ~ AGE + OBE,
  GHT ~ HT + MG
)

Complete DAG Visualization

complete_dag |>
    ggdag() +
    theme_dag()

DAGs and Scientific Complexity

So What?

I think it’s worth stepping back and considering how complex many scientific questions really are
We do not always observe everythign we want to, but with the help of DAGs we can:
- Identify the variables that are important to the question
- Identify what we can actually observe
- Identify what we cannot observe
- Assess the limitations we face with causal analysis

Next Steps

We are now going to work through seven applications/case studies, each addressing a different aspect of visualization.
In our first case study, we will examine how to visualize spatial variation.

Side Quest: Simulated Data

In examples/dag_sim_data.qmd we simulated data that encodes the DAG structure.
Let’s take a very brief glance at that file to see how we simulated the data.

Applications

Application 1: Choropleths

In applications_1_choropleth.qmd we create a choropleth map of the United States using the simulated data from above
We show how even with maps, small multiples help find clarity

Application 2: Forest Plots

In applications_2_forest_plots.qmd we create a forest plot of the data using the same simulated data.
We show how spatial variation can often be visualized better without actual maps.

Application 3: Ridgeline Plots

In applications_3_ridge_plots.qmd we create a ridgeline plot of the data using the same simulated data.
We show how ridgeline plots can show distributional differences across groups.

Application 4: Distributions

In applications_4_distributions.qmd we examine distributions of variables
Wecreate box plots and violin plots to show the distribution of a variable across a population.

Application 5: Time Trends

In applications_5_time_trends.qmd we examine time trends of variables
We create line plots to show how a variable changes over time.

Application 6: Scale Honesty & Effectiveness

In applications_6_scale_honesty.qmd we examine how axes can be used to mislead or clarify
We briefly discuss how what counts as dishonest may be contextual

Application 7: Visualizing Models

In applications_7_visualizing_models.qmd we visualize model outputs
We create coefficient plots and model predictions to help interpret the model

Summary and Takeaways

Visualization choices should be driven by the scientific question
Experimentation is key; often the “obvious” plot is not the best plot
Iterate: build in layers; get the basics down; then polish
Use small multiples, annotations, and thoughtful scales to guide the viewer